radeonsi: set the convergent attribute where needed

Reviewed-by: Dave Airlie <airlied@redhat.com>
This commit is contained in:
Marek Olšák 2017-02-26 19:00:44 +01:00
parent ef883fc554
commit d5d74fe2b5
2 changed files with 7 additions and 3 deletions

View File

@ -861,13 +861,17 @@ ac_build_ddxy(struct ac_llvm_context *ctx,
args[1] = val;
tl = ac_build_intrinsic(ctx,
"llvm.amdgcn.ds.bpermute", ctx->i32,
args, 2, AC_FUNC_ATTR_READNONE);
args, 2,
AC_FUNC_ATTR_READNONE |
AC_FUNC_ATTR_CONVERGENT);
args[0] = LLVMBuildMul(ctx->builder, trbl_tid,
LLVMConstInt(ctx->i32, 4, false), "");
trbl = ac_build_intrinsic(ctx,
"llvm.amdgcn.ds.bpermute", ctx->i32,
args, 2, AC_FUNC_ATTR_READNONE);
args, 2,
AC_FUNC_ATTR_READNONE |
AC_FUNC_ATTR_CONVERGENT);
} else {
LLVMValueRef store_ptr, load_ptr0, load_ptr1;

View File

@ -5207,7 +5207,7 @@ static void si_llvm_emit_barrier(const struct lp_build_tgsi_action *action,
lp_build_intrinsic(gallivm->builder,
HAVE_LLVM >= 0x0309 ? "llvm.amdgcn.s.barrier"
: "llvm.AMDGPU.barrier.local",
ctx->voidt, NULL, 0, 0);
ctx->voidt, NULL, 0, LP_FUNC_ATTR_CONVERGENT);
}
static const struct lp_build_tgsi_action tex_action = {